A Source Code Vulnerability Detection Method Based on Adaptive Graph Neural Networks

A Source Code Vulnerability Detection Method Based on Adaptive Graph Neural Networks

연구 분야: Strategies

학회: ASEW '24: Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering Workshops

This paper proposes a mobile application vulnerability detection method based on Code Property Graphs (CPG) and adaptive graph neural networks. The method first converts source code into CPGs, then uses CodeBERT to vectorize CPG nodes. Subsequently, high-level graph features are extracted through graph centrality analysis, and an adaptive graph neural network model combining Transformer's adaptive attention mechanism and Graph Convolutional Networks (GCN) is designed for feature learning and vulnerability detection. Experimental results show that this method achieves an F1 score of 82.9% on real vulnerability datasets, an improvement of 13.6%-49.9% compared to existing methods. Ablation experiments further validate the effectiveness of each key component. This research provides new insights and effective methods based on deep learning for mobile application security, demonstrating high application value and practical significance.